XI. SPEECH COMMUNICATION A. M. H. L. Hecker J. M. Heinz D. L. Hogan A. P. Paul Dr. T. T. Sandel Jane B. Arnold P. T. Brady O. Fujimura$ H. Fujisaki Prof. M. Hallet Prof. K. N. Stevens Prof. J. B. Dennis Dr. A. S. House SOME SYNTHESIS EXPERIMENTS ON STOP CONSONANTS IN THE INITIAL POSITION As a result of studies at Haskins Laboratories (1), and elsewhere (2, 3,4), it has been recognized that formant transitions constitute important cues for the recognition of initial voiced stop consonants. Some generally accepted rules concerning the formant transitions associated with different consonants have emerged from these studies, but there are considerable areas of disagreement. We have carried out a series of speech- synthesis experiments that are designed to supplement the results obtained previously and to study certain features that have not been examined in detail heretofore. Two different speech synthesizers have been used in the present experiments: dynamic analog of the vocal tract (called DAVO), a and a terminal analog (called POVO). The characteristics of the synthetic speech produced by DAVO are controlled in terms of the articulatory configuration (5). The control system of this synthesizer specifies precisely the time course of the articulatory changes, as well as the time-variant char- acteristics of the source excitation (buzz and/or noise); fundamental frequency of the buzz, by the buzz. that is, the amplitudes, the and the amount of amplitude modulation of the noise The same control system is used for POVO, but in this case the formant transitions are specified directly and the formant frequencies are changed in a piecewiselinear fashion. POVO can be excited by the same accurately controlled sources that are used to excite DAVO. The naturalness of the sounds generated by these synthesizers is generally very good when the variables are appropriately controlled, and it is often pos sible to generate stimuli that are judged almost unanimously to be members of a given phoneme. Consequently, we are in a position to evaluate not only the so-called primary cues, but also the secondary cues whose effects might be obscured if stimuli of poor quality were used (6). Before performing the detailed synthesis experiments, we made a series of informal evaluations of the sounds produced by POVO with a buzz source. We found that the rules This work was supported in part by the U.S. Air Force Command and Control Development Division under Contract AF19(604)-6102; and in part by National Science Foundation. On leave, 1960-61, as Guggenheim Fellow at the Center for Study of the Behavioral Sciences, Stanford University. SOn leave from the Research Institute of Communication Science, Electro-Communications, Tokyo, Japan. 153 University of (XI. SPEECH COMMUNICATION) provided by the Haskins group for the most favorable transitions of the second and third formants were generally appropriate, but the identifiability or the naturalness of the sounds was quite different for different stop-vowel combinations. Some syllables were ambiguous or not acceptable as the intended sound on an absolute-judgment basis, even when the control parameters were adjusted to give stimuli of optimum quality. The parameters that were controlled were the starting points and the duration of the linear shift of the lower three formants; the onset characteristics and the changes in frequency of the buzz source were also controlled. be generated with DAVO.) (In some cases more acceptable sounds could Synchronization of the time clock (5) to the buzz source was effective in improving the quality of initial voiced stops. Use of a very fast shift of formant frequencies was sometimes necessary, particularly for the study of the most favorable transition rate for a syllable /b/ + vowel. In this connection, we modified the output stage of the POVO circuits so that no appreciable thump would affect the output waveform even in a 10-msec transition (7). a. Voiced Stop Followed by /E/ The informal tests discussed above had demonstrated that reasonably satisfactory versions of the voiced stop consonants /b/, /d/, and /g/ can be obtained with both POVO and DAVO when these consonants are combined with the vowel /c/. Synthetic syllables with this vowel context were selected, therefore, as test materials for detailed studies of the influence of formant transition rate and of the effect of introducing an initial noise burst on the identifiability of voiced stop consonants. Stimuli for formal listening tests were prepared for POVO and DAVO. For the sounds generated by POVO, appropriate starting frequencies for the formants at the beginning of the transition were determined by listening to the POVO output. values finally selected and employed in the formal tests are given in Table XI-1. The The formant frequencies for the vowel were selected similarly by informal listening; the pre viously reported results for natural utterances (8) were taken into consideration. For DAVO an articulatory configuration appropriate for the vowel /E/ was determined by methods that took into account data from previous experiments with DAVO (5) and provided a perceptual quality, formant frequencies, with known data for the vowel. and a vocal-tract geometry consistent The articulatory configurations for the consonants (the starting points of the transition) were determined in similar fashion. frequencies for these DAVO configurations are listed in Table XI-1. Measured formant The frequencies of the first formants for the consonants were rather high because a sufficiently narrow constriction could not be attained with the continuously controlled circuit elements. In each of two similar experiments, one with POVO and another with DAVO, 96 test items were recorded in random order on a magnetic tape. for transition durations of 10, 20, 30, Four formant transition rates and 40 msec were provided for each of the 154 (XI. Table XI-1. SPEECH COMMUNICATION) Formant frequencies for the consonants (the starting points of the transitions) and for the vowel /E/ for the stimuli used in the tests of identification of voiced stops followed by /A/. Formant Frequencies (cps) b d g 1 185 185 200 490 F2 1150 1750 2100 1750 F3 2200 2700 2250 2300 F4 3300 3300 3300 3300 1 240 240 270 485 F2 1160 1740 2060 1725 F 3 2200 2750 2260 2300 F4 3200 3680 3420 3170 F POVO F DAVO syllables /bE/, /dE/, For POVO, half of the /g/-stimuli were produced with and /gE/. a short burst of noise at the onset of the syllable (gl), and the rest were produced without the noise (go). The noise burst was generated by mixing a short burst of white noise with the buzz signal at the input to the POVO circuit. For DAVO, a similar noise burst was applied at the glottis for half of the g-stimuli, and at the articulatory constriction for half of the d-stimuli. In one test (a forced-judgment test), 6 subjects were asked to identify each of the stimuli as one of the three syllables; and in another test (an unforcedjudgment test), as one of five possibilities, /bE/, The results are given in Table XI-2, /dE/, /gE/, /E/ or "something else." and from these data the following conclusions have been drawn. (a) For all of the three stops practically perfect identification could be obtained without the use of the noise burst, when the transition time was appropriate. This was true for sounds generated both by POVO and by DAVO. (b) The most favorable transition duration was relatively short for /b/, long for /g/, and intermediate for /d/. (Preliminary tests showed that a longer duration (as long as 60 msec) was acceptable for /g E/.) (c) There is considerable difference in the score when the forced judgment is com- pared with the unforced judgment. transition for /b/ In the POVO experiment, for example, a 10-msec and a 20-msec transition for /d/ tion (92 per cent and 94 per cent, one from among /b/, /d/, gave rise to almost perfect identifica- respectively) when the subjects were forced to choose and /g/, whereas for the unforced judgment the scores were 155 (XI. SPEECH COMMUNICATION) 62 per cent and 73 per cent, respectively. (d) In the DAVO experiment, a marked effect of the noise burst on the recognizability was observed. When the transition rate was relatively high, the presence of the burst helped to separate /g E/ from /E/. from both /g guishing /dE/ For a slower transition, it was helpful in distin- E/ and /E/. The use of a noise source thus expanded the range of the appropriate formant-transition rates. Table XI-2. Correct judgment score for the stimuli /bE/, /dE/, and /gE/ produced by POVO and DAVO. The subscript 1 to the consonants indicates the presence The consonants without of the noise burst; the subscript o, its absence. subscripts were produced without noise. Numbers represent the percentages of correct judgments out of 48 votes when there is no subscript, and out of 24 votes when there is a subscript, for the forced-judgment test. Numbers in parentheses represent the results of the unforced-judgment test. POVO Duration of Formant Transition b d 10 msec 92 (62) 44 (12) 50 (25) 54 (0) 20 msec 100 (100) 94 (73) 88 (58) 88 (62) 30 msec 100 (100) 98 (100) 100 (92) 100 (100) 40 msec 98 (100) 100 (100) 100 (96) 100 (96) g1 go DAVO Duration of Formant Transition b dI d gl go 10 msec 100 (100) 83 (83) 79 (79) 96 (100) 79 (54) 20 msec 100 (100) 92 (100) 100 (92) 100 (100) 92 (83) 30 msec 100 (92) 100 (92) 88 (67) 100 (100) 100 (92) 40 msec 100 (92) 88 (83) 92 (67) 100 (100) 100 (96) Considerable difference can be observed in the preferred transition time when the score obtained for DAVO is compared with that for POVO. For DAVO, the subjects always preferred faster transitions and, in particular, they could identify /b/ with a very fast transition (10-msec duration) without difficulty. The difference may be partially ascribable to a smoothing circuit (with a 5-msec time constant) which was 156 (XI. SPEECH COMMUNICATION) employed for the control signal of DAVO, but not for POVO. Another possible explana- tion is that the nonlinear transition of the formants (at the acoustic level), which depends automatically given by DAVO, can have a per- on the particular configuration and is ceptual effect that is significantly different from that of the linear formant transitions given by POVO. b. Cues for /ga/ In the informal tests described above the syllable /ga/ was one of the stop + vowel combinations that could not be synthesized successfully by appropriate selection of a linear formant transition. Further informal evaluations of this syllable as generated by the POVO synthesizer indicated that the quality of the output could be improved consider ably by modifying the syllable in the following manner: (a) by mixing a short burst of noise with the buzz excitation at the initial portion of the syllable; (b) by using a higher starting point for the first formant, for example, 450 cps instead of 200 cps; and (c) by delaying the start of the formant transition, thereby leaving a short buzz -excited portion with stationary formants. Formant frequencies for the consonants and for the vowel /a/ in the tests of cues for the syllable /ga/ Table XI-3. b d g g' a F1 200 200 200 450 750 F 2 800 1800 1500 1500 1100 F3 2000 2800 1800 1800 2500 F4 3300 3300 3300 3300 3300 Formant Frequencies (cps) Table XI-4. Percentages of votes indicating preferences for the stimulus over that of the column. For the stimuli represented by gl starting point of the first formant is at 225 cps; for g 3 450 cps. For gl and g 3 , the formant transition starts at the of the row and g 2 the and g 4 , at same time as the buzz onset; for g 2 and g 4 it starts 10 msec later. (See Fig. XI-1.) g1 g2 g2 31 g3 85 94 g4 94 92 157 g3 85 (XI. SPEECH COMMUNICATION) In order to evaluate the effects of these features, we have conducted two listening tests. In the first experiment, the effect of these features in combination was ascer- tained by an absolute-judgment test. POVO by a procedure and /gE/. Syllables /ba/, /do/, and /ga/ were produced by similar to that described above for the synthesis of /bE/, /d E/, The formant frequencies for the consonants (the starting points), for the vowel, are given in Table XI-3. quencies for /g/, we synthesized another /g a/ in which the three were included. This second /ga/, the three other syllables /ba/, tape for the listening test. as well as Using the same second- and third-formant freadditional features which will be called the g'-syllable, was mixed with /da/, and /ga/ in randomized order to make a recorded The time structures of the stimuli, as specified by the con- trol signals, are shown in Fig. XI-1. The control signals for b, d, and g are compared BUZZ FUNDAMENTAL FREQUENCY (100 CPS -- * 125 CPS) BUZZ AMPLITUDE r NOISE AMPLITUDE FORMANT TRANSITION -10 0 10 20 30 40 50 100 TIME (M SEC) Fig. XI-1. Forms of the control signals for the fundamental frequency of buzz source, buzz amplitude, noise amplitude, and formant frequencies. Solid lines are for the stimuli b, d, and g; broken lines, for the stimulus g'. with those for g'. four syllables. The recorded tape contained a total of 40 items, 10 for each of the Eleven subjects listened to the stimuli, and they were forced to choose one from among /ba/, /do/, 99 per cent of 110 votes, and /ga/. and as /d/ The results were that b was judged as /b/ by 1 per cent (one vote); d was judged correctly without any error (110 unanimous votes); g was judged as /d/ 19 per cent, and as /b/ by by 2 per cent; g' was judged as /g/ by 79 per cent, as /g/ by 99 per cent and as /d/ by by 1 per cent (one vote). The judgments of the g-items varied considerably from subject to subject, the number of /g/ responses ranging from 7 to 0 votes out of 10. 158 For most subjects, however, (XI. SPEECH COMMUNICATION) it is clear that the shift from /d a/ to /ga/ was caused by the three additional features. a second experiment In order to estimate the relative importance of the three cues, was conducted. By preliminary listening, it was found that an identifiable /ga/ was not obtained unless the first feature - the presence of noise - was included. Therefore, the remaining four combinations of presence or absence of the last two features were compared in a paired comparison test, with the noise always employed. Twenty-four pairs consisting of 12 different pair items were recorded in random order, and the subjects were asked to judge which of the pair was more like an utterance of /ga/. item was presented twice for one judgment. Eight subjects participated in the test, and five of them repeated the same test five months later. marized in Table XI-4. All of these judgments are sum- The designations of the four stimuli, gl' explained in Table XI-4. Every pair g2 g 3 , and g 4 , are A total of 52 judgments was made for each of the 6 combina- tions. Versions g 3 and g 4 were preferred to either gl or g 2 in 91 per cent of the judgments, and hence the higher location of the starting frequency of the first formant definitely improves the quality. The delay of the start of the formant transition was preferred when the other condition (concerning the first formant) was favorable (85 per cent preference for g 4 over g 3 ). The delay feature alone, on the other hand, does not seem to improve the sound. The stimulus designated as g 4 in this experiment is approximately identical to the g' in the previous absolute -judgment test. It is interesting to note that a noise of 10-msec duration preceding the voice onset was necessary to make an unambiguous /ga/. This duration of unvoiced hiss is often observed in spectrograms of natural utterances of /g/ syllable. in the initial position in an isolated It is plausible that the reaction (in the low-frequency region) of the vocal tract to the vocal cords is particularly appreciable in the case of the velar stop because the cavity behind the occlusion is very small. An appreciable disturbance or a delay of the onset of regular vibration of the vocal cords can be expected as a result of the occlusion or its sudden release. A higher location of the starting point of the first-formant transition can be a cue for velar stops as opposed to dental or labial stops. significantly higher for /g/ of /g/ than for /b/ or /d/ The locus of the first formant may be because the cavity behind the occlusion is substantially smaller than that for other stops, and possibly because the glottis is more open for /g/ than for /b/ or /d/. A value as high as 450 cps, however, does not necessarily represent the actual resonant frequency for the g-configuration. It is quite possible that the initial part of the rapid transition is overlooked in the perception of the natural utterance. In view of this, the preference for the delayed formant transition is particularly interesting. In the spectrograms of natural utterances of /ga/, a char- acteristic separation of the consecutively located second and third formants takes place (2, 3), but this separation starts very gradually and apparently only at a later stage 159 (XI. SPEECH COMMUNICATION) relative to the start of the first -formant transition. This characteristic is caused by a nonlinear relation between the resonant frequencies and the articulatory dimensions. Thus the entire pattern of the formant transition can be simulated effectively by the use of the additional features, if the initial part of the rapidly changing first formant is overlooked. c. Effect of a Fundamental Frequency Inflection on Tense-Lax Opposition This portion of the report describes a study of the cues for tense-lax opposition of stop consonants in the initial position, with the DAVO synthesizer used. consisted of the consonant pair g-k in combination with the vowel /8/. The test items Test stimuli were prepared with 10 different buzz onset times (in steps of 10 msec), and the presence absence of a fundamental frequency inflection was also controlled. The fundamental fre- or quency, when it had an inflection, was shifted linearly from 70 cps to 100 cps during the initial 150-msec period of the voicing. A white-noise source was applied at the glottis input of the electrical vocal tract during the initial portion of the syllable. started abruptly (with an onset time of I msec), The noise and then immediately decayed linearly to zero in 50 msec. The vocal-tract configurations for the consonant and the vowel were similar to those used in the previous experiment, and the g-configuration was shifted toward the 8 -configuration with a 40-msec transition time. For all test stimuli, the start of the noise and the configuration transition occurred simultaneously. The buzz started abruptly (with a 1-msec ramp) and the inflection started at the same time. In this experiment the ramp starting times were not synchronized with the buzz waveform. Eighty test items containing 4 each of 20 different stimuli were recorded in a randomized order. /gE/ The subjects were asked to decide whether the stimulus sounded like or like /k E/. the listening test. Ten American subjects and three Japanese subjects participated in The results for two subjects who gave particularly interesting responses are shown in Fig. XI-2. Both subjects P. B. and H. F. complete (4 votes) jump from g to k in one step, that is, showed the same with a 10-msec buzz onset time, when there was no fundamental frequency inflection. shift of the When the inflec - tion was present, this boundary shifted invariably to a later onset time for all subjects, but the curve of the stimuli with inflection present appeared considerably different from subject to subject. Subject P. B. did not give any g-responses if the onset time was 50 msec, or more, after the burst of noise; H. F. , on the other hand, was reluctant to give a k-vote, even with an unvoiced period as long as 80 msec, but he had no problem when there was no inflection. This subject (H. F.) is Japanese, jects did not show this marked behavior. but other Japanese sub- This tendency, however, American subjects, although usually less markedly. jects is shown in Fig. XI-3. 160 was often seen in The average for 10 American sub- -- WITHOUT INFLECTION OF FUNDAMENTAL WITH INFLECTION FREQUENCY FREQUENCY OF FUNDAMENTAL 100 °/ / ------ V k UJ V) z 0 H. F 0 Li Li ,N / P.B. i P. . HF. 0 < '7 0 u UJ >r Ii , -- - W1 /i L a 9 100% L 0 30 20 10 50 40 70 60 80 90 BUZZ ONST Fig. XI-2. as a function of the Responses /kE/ versus /gE/ two conditions of inflection of the fundamental are shown for two subjects, P. B. and H. F. sents the percentage of k-judgments (as opposed buzz onset time for Curves frequency. The ordinate repreto g-judgments). 1009 k Ui (n WITH INFLECTION OF FUNDAMENTAL / a FREQUENCY WITHOUT INFLECTION OF FUNDAMENTAL FREQUENCY --- 0. 0. UJ w 0 at I0 LJ Li 0Li a g 100, 0 iO 20 30 40 50 60 70 80 90 BUZZ ONSET Fig. XI-3. Averaged responses of 10 subjects for the same experiment as in Fig. XI-2. 161 (XI. SPEECH COMMUNICATION) It would appear that an inflection of the fundamental frequency, as used in this experiment, is not favorable for the naturalness of syllables consisting of a tense stop + a vowel, whereas it is favorable for voiced stops. Also, the judgment can possibly be switched from tense to lax just by changing the inflection, boundary region. if the buzz onset is in the For the initial stops of English, the primary cue of tense versus lax distinction is considered as a delay of buzz onset time associated with a certain amount of unvoiced hiss, while the inflection in the fundamental frequency of voice constitutes a secondary cue. A similar experiment has been conducted for the alveolar stops combined with the vowel /i/. In this experiment the noise was inserted at the articulatory constriction, and the ramp initiation was synchronized with the buzz. The range of buzz onset times was decreased, and 10 judgments were collected from each subject for each condition of the test stimuli. The result was very similar to that of the experiment just described, except that the shift from /di/ to /ti/ occurred at an earlier value of the buzz onset time. O. Fujimura References 1. A. M. Liberman, P. C. Delattre, F. S. Cooper, and L. J. Gerstman, The role of consonant-vowel transitions in the perception of stop and nasal consonants, Psychol. Monogr. Vol. 68, No. 8, pp. 1-13, 1954; K. S. Harris, H. S. Hoffman, A. M. Liberman, P. C. Delattre, and F. S. Cooper, Effect of third-formant transition on the perception of the voiced stop consonants, J. Acoust. Soc. Am. 30, 122-126 (1958); H. S. Hoffman, Study of some cues in the perception of the voiced stop consonants, J. Acoust. Soc. Am. 30, 1035-1041 (1958); P. C. Delattre, A. M. Liberman, and F. S. Cooper, Acoustic loci and transitional cues for consonants, J. Acoust. Soc. Am. 27, 769-773 (1955). 2. M. Halle, G. W. Hughes, and J. P. Radley, Acoustic properties of stop consonants, J. Acoust. Soc. Am. 29, 107-116 (1952). 3. E. Fischer-J rgensen, Phonetica 2, 42-59 (1954). Acoustic analysis of stop consonants, Miscellanea 4. C. G. M. Fant, Acoustic Theory of Speech Production (Mouton and Company, The Hague, 1960). 5. G. Rosen, Dynamic Analog Speech Synthesizer, Technical Report 353, Research Laboratory of Electronics, M.I.T., Feb. 10, 1960. 6. A. S. House, Formant bandwidth and vowel preference, Quarterly Progress Report No. 56, Research Laboratory of Electronics, M. I. T., Jan. 15, 1960, pp. 172-173. 7. Final Report on Research on Speech Synthesis, December 1959. J. Report AFCRC-TR-60-101, 8. G. E. Peterson and L. L. Barney, Control methods used in a study of the vowels, Acoust. Soc. Am. 24, 175-184 (1952). 162